Classifications

• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
  • G06Q10/00—Administration; Management
  • G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
• • G—PHYSICS
  • G06—COMPUTING; CALCULATING OR COUNTING
  • G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
  • G06N7/00—Computing arrangements based on specific mathematical models
  • G06N7/01—Probabilistic graphical models, e.g. probabilistic networks

Definitions

• the present invention addresses the prediction of malfunctions of components, units, and fleets of technical entities.
• the invention is in particular applicable to industrial asset management, but also to other fields.
• US8200600 disclose the use of Markov models for predicting certain parameter values or states of the components. These models have the advantage that they can be used for a variety of component types.
• One aim of the invention is to find a method for predicting a malfunction of a technical entity, where the method is independent of component types and can compute the probability of this malfunction for any future point in time.
• Another aim of the invention is to find a computationally- efficient method for predicting a malfunction of a component and/or unit.
• Another aim of the invention is to find a method that reduces the model creation and configuration effort required to predict a
• Another aim of the invention is to find a method to configure such a method/apparatus for predicting a malfunction of a component/unit of a technical entity.
• Another aim of the invention is to find a method for predicting a malfunction of a component based on the component's available condition and process parameters.
• the invention achieves these aims through a method, computer program or apparatus for predicting component malfunctions comprised of the following steps:
• determining a conditional probability for a malfunction at the future point in time given the current value of the parameter based on the conditional probability distribution of the parameter for the future point in time given the current value of the parameter and on a conditional probability distribution for the malfunction given the parameter.
• the probability of the malfunction is calculated in a second step on the basis of the conditional probability of the malfunction given the parameter.
• the invention achieves these aims through a method, computer program or apparatus for predicting a malfunction of a component, comprised of the following steps:
• transition matrix for each parameter on the basis of the values of this parameter, wherein the transition matrix comprises for the plurality of discrete value states of each parameter the probabilities to switch from one of the discrete value states to another of the discrete value states within a certain time period;
• This method for configuring the prediction method allows to determining the necessary probability distributions which is not trivial.
• conditional probability distribution for the malfunction given the parameter is determined based on the
• the component is a mechanical or electrical component.
• the malfunction of the component is predicted on the basis of the conditional probability for a malfunction at the future point in time given the current value of the parameter. This could be achieved e.g. by outputting or displaying the resulting conditional probability for a malfunction at the future point in time given the current value of the parameter or by further processing the resulting conditional probability for a malfunction at the future point in time given the current value of the parameter.
• the current value of the parameter of the component is measured.
• this measurement is performed at the component, e.g. by a sensor.
• the steps for determining the conditional probability for a malfunction at the future point in time given the current value of the parameter are performed at a server.
• This server is preferably remote from the component or the components or at least some of the components.
• the current value is transmitted over a communication network to the server.
• the method comprises the further steps of:
• transition matrix wherein the transition matrix is based on probabilities to switch from one of a number of discrete value states to another of the discrete value states;
• conditional probability distribution for the malfunction given the parameter being a conditional probability
• step of determining the conditional probability distribution of the parameter of the component for the future point in time given the current value of the parameter comprises the step of determining a conditional probability distribution of the discrete value states of said parameter for the future point in time given the current discrete value state of the parameter based on the current discrete value state of the parameter and on the transitional matrix;
• the step of determining the conditional probability for the malfunction at the future point in time given the current value of the parameter comprises the step of determining a conditional probability for the malfunction at the future point in time given the current value state based on the conditional probability distribution of the discrete value states of said parameter for the future point in time given the current value state and on the conditional probability distribution for the malfunction given the discrete value states of said parameter.
• providing shall have the meaning inputting.
• Inputting information comprises both, inputting the information directly or inputting another information from which the mentioned information is retrievable, e.g. by calculation.
• providing means calculating the information.
• the intervals between the discrete value states are equidistant.
• the intervals between the discrete value states are logarithmic or exponential. [0029] In one embodiment, the size of a value interval corresponding to a discrete value state of the discrete value states depends on the
• the method comprises the step of recording the values of the parameter and the step of determining the transition matrix on the basis of the recorded values of the parameter.
• the method comprises the step of
• the transition matrix is based on the probabilities to switch from one of a number of discrete value states to another of the discrete value states within a determined time period and the probability of the malfunction given the current discrete value state of the parameter is determined for a number of future points in time, wherein the future points in time are multiples of the determined time period.
• the method comprises the step of recording the values of the parameter.
• the statistical significance of the data underlying the probability for the malfunction at the future point in time given the current value of the parameter is calculated based on the statistical significance of the data underlying the conditional probability distribution of the parameter given the current value of the parameter.
• the probability for the malfunction given the current value of the parameter is determined for a number of future points in time.
• the probability distribution for the malfunction given a number of future points in time is displayed for the future points in time together with a statistical significance of the data underlying the probabilities for the malfunction.
• the method comprises the step of estimating a remaining useful life of the component on the basis of probabilities for the malfunction of the component given the number of future points in time.
• a ) ] is based on the integral I P t o+ At ( a
• the component is a gas turbine.
• the malfunction component is one of a bearing defect, compressor defect, combustion defect, flow malfunction, turbine malfunction, or output malfunction
• the parameter for predicting the malfunction is one or a combination of temperature, lubricant condition in the bearings or in the oil tank, shaft or casing vibration, temperature, flow and pressure of the gas in the turbine, speed of rotation, electric output of the coupled generator, ambient air temperature or humidity.
• the component is a transformer.
• the malfunction of the component is one of an insulation defect, cooling system defect or oil aging, and the parameter for predicting the
• the component is a diesel engine
• the malfunction of the component is one of bearing defect, gearbox tooth wear, or turbo charger defect
• the parameter for predicting the malfunction is temperature, vibrations, lubricant condition, speed, or a combination thereof.
• the unit is an industrial asset.
• the method comprises the step of
• the configuration method comprises the step of providing probabilities for the single malfunctions of each component comprising:
• the relative probabilities are each indicated by a scale between a minimum and a maximum and the absolute probabilities of the single malfunctions are determined by rescaling the relative probabilities so that the sum of all relative probabilities yields the probability for having a malfunction of the corresponding component.
• the configuration method for each single malfunction of the corresponding component additionally the relative impact of the single malfunction is received.
• the step of providing the conditional probabilities of the discrete value states of the parameters given the corresponding single malfunction comprises for each single malfunction of each component:
• the number of parameter intervals is smaller than or equal to five, preferably is equal to four.
• the method of the configuration comprises the step of determining a wear measuring quantity for each component.
• Fig. 1 shows an embodiment of a method for predicting a single malfunction of a component
• Fig. 2 shows an embodiment of a method for predicting a malfunction of a unit
• Fig. 3 shows an embodiment of an apparatus for predicting a single malfunction of a component
• Fig. 4 shows an embodiment of an apparatus for predicting a malfunction of a unit
• Fig. 5 shows schematically a steam turbine as an exemplary embodiment of a component, with its subcomponents and monitored parameters
• Fig. 6a shows an example diagram for presenting the single malfunction probabilities of a component, for different single malfunctions and different points in time;
• Fig. 6b shows an example diagram for presenting the total probabilities of a malfunction of the components of the unit for different points in time
• Fig. 7 shows an embodiment of a method for configuring an apparatus for the prediction of a malfunction of a unit
• Fig. 8 shows an exemplary diagram for determining the probabilities of the single malfunctions of a component
• Fig. 9 shows an exemplary diagram for determining the conditional probabilities of the discrete parameter value states given a set of malfunctions.
• Fig. 10 shows an embodiment of an apparatus for configuring ar apparatus for the prediction of a malfunction of a unit.
• Fig. 10 shows an embodiment of an apparatus for configuring ar apparatus for the prediction of a malfunction of a unit.
• a component is a subpart of a unit.
• the component is a functional subpart of the unit.
• the component is for example a wear part of the unit.
• the component could be a mechanical and/or electrical component of the unit for providing a corresponding mechanical and/or electrical function for the unit.
• a unit comprises a plurality of components providing together a certain functionality of the unit.
• the unit comprises different types of components.
• the unit is an industrial asset.
• An example for an industrial asset unit is a power plant with components such as a nuclear reactor, a gas turbine, a steam turbine, a power generator, a cooling system, a transformer, etc.
• Another example for an industrial asset unit is an upstream oil & gas production site with components such as a pump, a compressor, a gas turbine and/or internal combustion engine, a power generator, and a pipe system.
• Another example for an industrial asset unit is a cement plant with components like a raw mill, coal mill, rotary kiln, a clinker, an air-to-air cooler, etc.
• an industrial asset unit is a heavy truck or heavy mobile machine, such as those used in mining, with components such as a diesel engine, gearbox, hydraulic system, tools, etc.
• Another example for an industrial asset unit is a plane with components such as jet engines, fuselage, air brakes, doors, etc.
• An additional example for an industrial asset unit is a locomotive with components such as a diesel engine or electric motor, a gearbox, bogies with wheels, axles and a brake system, etc.
• a further example for an industrial asset unit is a transmission facility with transmission and/or distribution transformers, switches, etc. These examples shall not be exclusive. It is also possible to apply this method to technical entities that are non-industrial units, for instance hospitals with components such as medical diagnostic and/or surgical instruments. These include CT-tomographs, surgery robots, MRT-tomographs, etc.
• a fleet is an asset operator's collection of units.
• the asset operator can be the provider of a service or the manufacturer of a product, for instance an airline operating a fleet of planes or an electricity provider with a fleet of power plants.
• a transmission system operator with a fleet of power transmission equipment - transformers, switches, etc. - a mining company operating a fleet of mining trucks and mining tools, or an oil & gas company is operating a fleet of oil & gas facilities, etc.
• malfunction can be interpreted as operating conditions, mechanical, electrical, chemical or other problems that - if not treated - may cause degradation in performance, an unplanned shutdown, or a catastrophic failure.
• malfunction and type of malfunction are synonymous.
• single malfunction refers to one specific type of malfunction of the component or unit, while the total malfunction refers to any malfunction of the component or unit.
• probability of the malfunction can either be the probability of having such a malfunction or of not having such a
• the state-of-the-art parametric model is replaced by a stochastic model which can be used for any type of component.
• a stochastic model is presented and the steps for an embodiment of a method for predicting a single malfunction M k of a component is described and shown in Fig. 1 .
• At least one parameter is chosen for detecting the single malfunction M k of the component.
• the parameter is a physical, electrical, chemical, or other parameter of the component, such as vibration, temperature, water condition, oil condition, current, voltage, pressure, etc.
• only one parameter may be chosen, while in another embodiment at least two parameters may be chosen.
• the value states may be distributed equidistantly over the parameter value range, while in another embodiment the value states may be distributed logarithmically over the parameter value range. In yet another instance, the value states are distributed exponentially over the parameter value range. In another embodiment, the value states are distributed in dependence of the probability of the parameter. This means that value states including parameter values appearing with high probability encompass a smaller parameter value range than value states including parameter values appearing with lower probability. Therefore, it can be guaranteed that all value states contain a reasonable number of parameter values to create significant results. In the case that two parameters are chosen for detecting one single malfunction, the two- dimensional parameter value state space has to be divided into a number of discrete value states, with each value state covering a certain area in the parameter value state space.
• n- dimensional value state space would be divided in a number of discrete subspaces forming the discrete states.
• An alternative is to detect the single malfunction individually on the basis of each parameter and then to combine the parameter-specific conditional probabilities of the
• a transition matrix T for the selected parameter has to be provided with th entry in the i-th row and the j-th column:
• the transition matrix T is received. In another embodiment, the transition matrix T is computed based on the past values of the parameter of only this component. In yet another embodiment, the transition matrix is computed by the past values of the parameter of identical or similar components. In another embodiment, the transition matrix T is obtained through expert assessment. This embodiment is in particular important, when there are not yet data available in some or all of the parameter value states. In one embodiment, the transition matrix is updated periodically on the basis of new values of the parameter. All five embodiments can be combined or sub-combinations of them can be formed. The parameter values are typically recorded or measured during operation.
• the recorded values are cleaned before their usage in determining the transition matrix, e.g. by removing parameter values recorded during periods the component and/or the unit is not operated in a representative operating state and/or is not operated in a specific operating state.
• T or the transposed matrix T T is used.
• this probability distribution is received, e.g. in a computer, server, database, or other apparatus performing the method.
• this probability distribution is determined.
• C, ) could be statistically inferred through very disciplined condition monitoring and recording of Q together with empirical malfunction tests on M k . This would typically involve crash testing, or at least serious risk of asset failure and downtime, which in practice is rarely possible due to the high downtime costs.
• M k ) that the parameter is in a certain state C, when the malfunction M k occurs can be determined on the basis of expert assessments.
• the probability P( M k ) of the single malfunction M k can be retrieved in different ways. In one embodiment, it is determined on the basis of malfunction statistics for the component. In another embodiment, it is determined on the basis of malfunction statistics for identical or equivalent components. In another embodiment, it is determined on the basis of expert assessments as explained later with respect to Fig. 8.
• the probability P ( C, ) of the discrete state C can be computed from the data history of the parameter. The probability P ( C, ) of the discrete state C, can be periodically updated on the basis of new values/measurements of the parameter.
• the current value state C(t 0 ) of the selected parameters is provided.
• the state vector C(t 0 ) e R N and C eR N is a sparse vector comprising a " 1 " at the position referring to the current value state Ci and a "0" for all the other states Cj with i ⁇ j, representing the probability distribution.
• the current value state can be provided in any other form.
• the current value state C(to) is determined on the basis of a current measurement of the parameter value.
• the current value state C(t 0 ) is determined on the basis of an assessment or estimate of the parameter value.
• the current time is referred with to.
• the term "current time” in this invention is interpreted as the time of the last value/measurement of the parameter(s) available. This could be in one example the actual time for online monitoring or an hour ago or yesterday in another.
• a fourth step S4 the state C(t 0 +L*At) of the parameter at a future point in time t 0 + L*At is predicted.
• the time step At is a day, such that to corresponds to today and t 0 +L*At to L days later.
• any other time step is possible. This is performed on the basis of a stochastic process model, such as a Markov chain, by multiplying the current state C(t 0 ) with the transition matrix T
• step S5 the probability of a single malfunction Mk at the future point in time to+L*At is determined.
• the stochastic model combines Markov chains for predicting the discrete value state of the parameter at a future point in time to+L*At with the conditional probability distribution P( Mk I Ci ) for the malfunction Mk given the discrete states C, of said
• C ) e R N contains the N probability elements P( Mk
• C, ) for the malfunction Mk given each state Ci, for i 1 N. This allows the same stochastic approach for modelling malfunctions of any unit or component. There is no need to model the component to predict malfunctions.
• the shown order of the steps S1 to S5 is not mandatory; in particular S1 to S3 could be performed in any order.
• the method described before is performed by an apparatus for predicting the malfunction.
• An embodiment of such an apparatus is described in Fig. 3.
• Such an apparatus can be a server, preferably a server being remote from the component, a normal computer, a server computer or also any circuit adapted to perform the described method.
• Said apparatus could receive all the necessary information for the configuration and/or the current values of the parameter (or directly the current value states of the parameter) by an interface.
• the interface is connected to a communication network to receive the information for the configuration and/or the current values of the parameter from a remote location.
• the communication network or connection could be wired (e.g. Internet, LAN, or combination of those) or wireless (WLAN, Near-field communication, mobile telephone
• the prediction for the malfunction given the new current value is determined as described above and a certain output is prepared to show the prediction of the malfunction, e.g. in a visual way.
• the current value can be sent periodically to the apparatus (e.g. every hour, every day, every week).
• the apparatus for performing the method mentioned above is part of a system including also the component, a sensor for measuring the current value of the parameter of the component and a communication network connecting the sensor with the apparatus for performing the method mentioned above.
• the prediction of the malfunction of the component can be fully automated by measuring the current value of the parameter by the sensor, transmitting the parameter to the apparatus and predicting the malfunction by the apparatus as described above.
• the current value of the parameter could be determined or measured by a human and the measurement result could be inserted in an input interface of said apparatus or an input interface transmitting the measurement result, i.e. the current value, to the server.
• Fig. 2 shows an instance of a method for predicting a
• a first step S1 1 for each component I and for each malfunction M
• the probabilities of all single malfunctions of a component can be combined to the total probability of malfunction of the corresponding component. The combination can be based on stochastic rules considering their degree of dependence. For example, in the case of independent malfunctions Mik the probability of the single malfunction P(M
• the probabilities for different single malfunctions can be combined to a total probability of a malfunction of the component considering the different effects of all single malfunctions. If for some or all components only one single malfunction is determined, the probability of this one malfunction corresponds to the total probability of malfunction of the component.
• the probability for a malfunction of the unit is determined on the basis of the total probabilities of malfunction of the components of the unit. In one embodiment, the probability of
• the steps of the method described in relation to Fig. 2 can be performed by the same apparatus performing the method described in Fig. 1 .
• An embodiment of such an apparatus is described in Fig. 4.
• another apparatus is used or that at least two apparatuses are used to predict the malfunction of the plurality of components and to combine those probabilities of malfunction of the single components in one of those apparatuses or in another apparatus.
• the intermediate server could receive the current values automatically from a sensor at the components or by interfaces which allow users to insert the measurement results or combinations of both.
• Fig. 3 shows an embodiment of an apparatus 10 for determining the probability of a single malfunction M k of a component.
• the apparatus 10 comprises a transition matrix section 1 1, a current state section 12, a conditional probability section 13, a parameter value prediction section 14, and a component malfunction prediction section 1 5.
• the transition matrix section 1 1 is configured to provide the transition matrix T.
• the transition matrix T is
• the transition matrix section 1 1 is configured to determine the transition matrix T on the basis of past measurement of the parameter of the component and/or of a comparable component.
• the current state section 12 is configured to provide the current value state C of the parameter used for predicting the malfunction M k . In one embodiment, the current state section 12 receives directly the current state C. In another embodiment, the current state section 12 is configured to determine the current value state on the basis of an empirical
• the conditional probability section 1 3 is configured to provide the conditional probability distribution P ( Mk
• the parameter value prediction section 14 is configured to predict the probability distribution of the discrete states C j (t 0 +L*At) at a future point in time to+L*At based on the transition matrix T and the current state vector C(t 0 ).
• the component malfunction prediction section 1 5 is configured to predict the probability of the single malfunction Mk in the future t 0 +L*At given the current value state C(t 0 ) based on the probability P( Ci(t 0 +L*At)
• C(to) ) of the state Ci(t 0 +L*At) at the future point in time t 0 +L*At with i 1 N given the current value state C(t 0 ) and the conditional probability distribution P ( Mk
• G ) for the malfunction Mk given the discrete value states Ci of said parameter with i 1 , N.
• the apparatus 10 computes in this embodiment the probability of the single malfunction Mk at the future point in time to+L*At given the current value state C(t 0 ) of the parameter.
• the apparatus 10 can determine the probability of the single malfunction Mk at the future point in time t 0 +L*At given the current value state C(t 0 ) of the parameter for different future points in time, i.e. for different L out of the set of natural numbers. This allows displaying a function for the probability of a malfunction over a future time horizon.
• a potential malfunction event or remaining useful lifespan can be estimated.
• the apparatus 10 is further configured to compute the significance of the data underlying the probability of the single malfunction M k in the future point of time t 0 +L*At given the current value state C(t 0 ) of the parameter for different future points in time.
• Fig. 4 shows an embodiment of an apparatus 20 for predicting the malfunction of a unit with a plurality of components.
• the apparatus 20 comprises the component total malfunction prediction sections 21 .1 to 21 . m and a unit malfunction prediction section 22.
• malfunction prediction section 10.ji is configured to predict the i-th single malfunction of the j-th component of the unit.
• each single malfunction prediction section 10.ji corresponds to the apparatus 10 shown in Fig. 3.
• Each component total malfunction prediction section 20.j is configured to predict the total probability of a malfunction of the corresponding j-th component on the basis of the probabilities of the nj single malfunctions of the j-th component.
• the unit malfunction prediction section 22 is configured to predict the malfunction of the unit on the basis of the m predicted probabilities of the malfunction of the m components of the unit. Therefor, the unit malfunction prediction section 22 computes the total probability of malfunction of the unit given the current parameter value states on the basis of the m total probabilities of the malfunction of the m components of the unit.
• this method has the advantage that it can be applied to every kind of component of a technical entity unit and to every kind of malfunction. Therefore, the malfunctions of all components can be computed by the same procedure. This allows a very easy scaling of the method from a small component to a fleet of technical entities.
• the component is a gas turbine.
• the gas turbine is coupled with a generator with a power output larger than 100 MVA, in particular larger than 1 50 MVA.
• Typical single malfunctions for a gas turbine are bearing defects, compressor defects, flow malfunction, turbine malfunction or output malfunction. One or any combination of the following parameters can be used for
• detecting/predicting the respective single malfunctions temperature, lubricant condition in the bearings or in the oil tank, and shaft or casing vibration for detecting/predicting bearing defects; temperature and pressure of the gas in the turbine for detecting/predicting flow problems; temperature, rotation speed of the turbine, and electric output of the generator for detecting/predicting turbine problems; ambient air condition and humidity for detecting/predicting output problems of the turbine.
• the component is a transformer.
• the transformer is configured for a power larger than 100 MVA, in particular larger than 1 50 MVA.
• the transformer has coils which are surrounded by oil.
• the transformer has cooling fans. Typical single malfunctions for a transformer are insulation defects or cooling system defects. One or any combination of the following parameters can be used for
• the component is a diesel engine.
• the diesel engine has a power of more than 200 kW, in particular more than 300 kW.
• Typical single malfunctions for a diesel engine are bearing defects or turbo charger defects, in particular a defect of the compressor of the turbo charger.
• FIG. 5 shows now a steam turbine 30 as an embodiment of a component.
• This steam turbine 30 comprises a turbine 31, a generator 32, an axle shaft 33, bearings 34.1 , 34.2 and 34.3 and a lube oil circuit 35.
• the turbine 31 and the generator 32 are mounted on the axle shaft 33 which is born by the bearings 34.1, 34.2 and 34.3.
• the bearings 34.1 , 34.2 and 34.3 are connected to the oil circuit 35 for exchanging the oil in the bearings 34.1 , 34.2 and 34.3.
• the oil circuit 35 comprises an oil tank 36 and an oil pump 37 which pumps lube oil from the tank 36 to the bearings 34.1 , 34.2 and 34.3 over the input link 38.
• the lube film bearings 34.1 , 34.2 and 34.3 are connected with the tank 36 via the oil pipe 39.
• the oil pipe 39 comprises the single drain pipes 39.1, 39.2 and 39.3 in which the lube oil coming from the bearings 34.1 , 34.2 and 34.3 is not yet mixed.
• the following parameters of this steam turbine 30 are measured: the vibrations of the axle shaft 33 at bearings 34.1 (V1 ), 34.2 (V2) and 34.3 (V3), the thrust vibration at bearing 34.3 (V4) and the casing vibrations of the generator 32 (V5); the temperature of the three bearings 34.1 (T1 ), 34.2 (T2) and 34.3 (T3); the lubricant condition in the three individual drain pipes 39.1 (L1 ), 39.2 (L2) and 39.3 (L3) and the lubricant condition in the tank (L4); and the electric output E1 of the generator 32 (E1 ).
• Fig. 6a shows a matrix of the probabilities of different single malfunctions for different future points in time.
• the malfunctions M1 to M12 are listed in rows and the future points in time in columns.
• the total probability of a malfunction for the component is shown as well in the last row. This total probability of a malfunction of the component can be determined on the basis of the single malfunction probabilities. Note that the computation of the probability of some malfunctions, here M1 1 and M12, were deselected and are thus empty. In this case, the malfunction probability of all malfunctions M1 to M 10 stays below 10% until April 4. Then, the prediction for April 1 1 shows an increased probability of malfunction M7, i.e. a particulate contamination, with more than 10%. If the next maintenance is before this date, this should be no problem.
• the table of Fig. 6a further shows the significance of the data underlying the calculated probabilities.
• a reduced significance i.e. medium predictive strength
• a low significance i.e. low predictive strength
• the significance of the data underlying the calculated probability shows already for the March 14 a low predictive strength.
• Fig. 6b shows now the total probability of each component of the unit, i.e. here of each component of a power plant analogue to Fig. 6a.
• the probabilities are written explicitly. It could be more intuitive to show the probabilities by discrete probability ranges, e.g. the parameter ranges below 10%, between 10% and 25% and above 25%. Those ranges could be shown by different colours and provides an intuitive way to show this information.
• Fig. 7 shows now a method for configuring the prediction apparatus/methods described above. In a first step S21 , the components of the unit are defined.
• a wear measuring quantity is provided with the components which allows estimating the remaining useful lifespan in general. This remaining useful lifespan is shortened, if the above described methods predict an earlier malfunction. It can be prolonged by removing the cause of the malfunction through maintenance, change of operational scenario, etc.
• step S22 the single malfunctions of each of the components are defined. This can be performed by a computer-based configuration tool, or by hand, or be received from a third party. In one embodiment, for each malfunction-component couple, the potential parameters for detecting such a single malfunction are defined as well. This facilitates later determination of the best parameters for malfunction prediction.
• step S23 the probability P(M j i)of each malfunction i for each component j is provided.
• the probability of each malfunction for each component can be computed, assessed, or received from a third party. This can be done by computer-based tool or by hand. For some components, there is historical data on all malfunctions that have occurred on this component or on equivalent components, in which case the probabilities can be estimated on the basis of this historical data. However, often such data is not available. Therefore, in another embodiment, these
• probabilities are determined by expert assessments. Since it is hard to tell the absolute probability of a single malfunction of a component, the input of the relative probability of this malfunction compared to the other malfunctions is asked. This relative probability can be expressed for example by a scale from 1 to 10. Since the expert knows from his
• FIG. 8 shows for the example of the steam turbine from Fig. 5 and Fig. 6a the twelve malfunctions and their relative likelihood. Assuming now that all malfunctions listed cover only a
• the relative impact of the corresponding malfunction can be related to the probability of a malfunction.
• the user will see the difference between the relative likelihood and the relative impact of a malfunction, which are often mixed. Therefore, the information about the relative impact improves the quality of the input about the relative likelihood.
• the value of 80% for the probability A of all considered single malfunctions for one component is in most cases a good value. But the probability of all considered single malfunctions for one component could also be input for the computation of absolute probabilities of the single malfunctions.
• step S24 the parameters of the components used for
• predicting the single malfunctions of the components are determined. This could be performed for example by selecting them from the available parameters of the components. However, this step could also be done implicitly in step S28.
• step S25 the data histories of the parameters from step S24 are obtained. Normally, this data for the parameters comes from the corresponding component itself. However, if the parameter values have not been recorded in the past, the data of a comparable component can be used.
• each parameter used for detecting/predicting a malfunction of a component is divided in a plurality of discrete value states.
• the exact number of discrete value states depends on the volume of data that can be obtained - the sampling time span and frequency. For example, the number of discrete states can be smaller than 50 or smaller than 40 or between 20 and 40.
• the size of the parameter value intervals of the discrete value states is equal. In another embodiment, the size of the parameter value intervals of the discrete value states is
• the size of the parameter value intervals of the discrete value states is exponential.
• the size of the parameter intervals of the discrete value states depends on the frequency of the measurements.
• the size of the parameter intervals of the discrete value states depends on the frequency of the measurements.
• step S27 the transition matrix is determined for each parameter.
• the dimension of the transition matrix is quadratic and the number of columns or rows corresponds to the number of discrete value states.
• the entries correspond to the probability to switch from one value state to the other.
• the transition matrices can be calculated on the basis of the measurements of the parameters or - as previously described - on the basis of measurements of the parameters of comparable components or on the basis of expert assessments.
• step S28 the probability distribution of parameter values over the discrete value states is determined.
• this probability distribution can be estimated on the basis of the historical parameter values.
• the distribution is taken from comparable components. In yet another embodiment, it is assessed by experts.
• step S29 the conditional probability distribution for a discrete value state of a parameter given a malfunction is provided for all parameters and malfunctions of the components.
• the parameter values of the corresponding components at the point of malfunction could be determined, and the probability distribution P(C
• M) is determined on the basis of data from a comparable component.
• M) is determined by expert assessments. Therefore, the parameter value range of each parameter is divided into a reasonable number of parameter intervals. Preferably the reasonable number is smaller than the number of states for this parameter.
• the reasonable number is smaller than ten, preferably smaller than six.
• Fig. 9 an example for such an expert assessment is shown for only two parameters of a first component and three malfunctions.
• the parameter value range of each parameter is divided into four value ranges.
• a green range wherein the parameter is expected to be during normal operation.
• a yellow range which is an alert range and covers more or less only outliers.
• An orange range which would raise an alarm, and a red range where an immediate shutdown of the component/unit is expected.
• the colours are only arbitrary and the ranges can also be named differently, for example first to n-th.
• the expert can now determine the probability of a parameter being in one of the four ranges given a certain malfunction.
• the parameter P1 .1 Since the parameter P1.1 is not influenced by the malfunctions M1 and M2, the parameter P1 .1 is expected to stay in the green range or at least shows the unconditional distribution of parameter values. However, the parameter P1 .1 is influenced by a malfunction M3. An expert knows by experience more or less how the parameter values are distributed over those four parameter value ranges/intervals. The information is then translated to the probability of the discrete value states given the single malfunction. This simplification allows the necessary information to be obtained from the expert.
• step S30 the conditional probabilities P(M
• Q ) P(Q
• Fig. 10 now shows an apparatus 40 for configuring a prediction apparatus for the malfunction of a unit.
• the apparatus 40 comprises a receiving section 41 , a dividing section 42, a transition matrix section 43, a first conditional probability section 44, a discrete value state probability section 45, a malfunction probability section 46, and a second conditional probability section 47.
• the receiving section 41 is configured to receive information about:
• the dividing section 42 is configured to perform step S26.
• the transition matrix section 43 is configured to determine the transition matrices.
• the first conditional probability section 44 is configured to determine the conditional probability P(C
• the discrete value state probability section 45 is configured to determine the probability P(C) of the discrete value states.
• malfunction probability section 46 is configured to determine the probability P(M)of the single malfunctions.
• the second conditional probability section 47 is configured to determine the conditional
• the method/apparatus for predicting a malfunction of a unit/component can be initiated. It is also possible to use the continuously received current parameter values to update T and P(C). It is also possible to use data which registers the single malfunctions of the components to update P(M k ). With the continuous value recordings of the parameters and the identified malfunctions related to the parameter values, P(C,
• the shown embodiments are only exemplary and shall not restrict the invention. All embodiments covered by the scope of the set of claims shall be included in the invention.
• the inventive step of determining a conditional probability distribution of a parameter of the component for a future point in time given the current value of the parameter was performed in the described embodiment by a discrete-time Markov chain (DTMC) model.
• this step can also be performed by a continuous-time Markov chain (CTMC) model. This can be performed by reducing the predetermined time period At.
• CTMC continuous-time Markov chain
• the formulas of the DTMC model can be replaced by the formulas for the CTMC model for computing the continuous transition probabilities.
• the parameter value state vector C(t) is computed by
• the transition matrix comprises the transition rates
• the CTMC model could be interesting for parameters with high-frequency or for near continuous online monitoring.
• the inventive step of determining a conditional probability distribution of a parameter of the component for a future point in time given the current value of the parameter was performed in the described embodiment by a Markov chain model dividing the parameter space into discrete parameter value states.
• other methods can also be used to determine the conditional probability distribution of a parameter of the component for a future point in time given the current value of the parameter.
• the distribution equation dP(a(t))/dt D(a) d 2 P(a(t))/da 2 with the probability distribution P(a(t)) of the parameter a at the time t and the distribution coefficient D(a) can be used to determine the conditional probability distribution P(a(t)
• the distribution coefficient could depend on the parameter a, or could be constant. For constant diffusion coefficients D the solution would be
• the invention is described here with respect to malfunctions of technical entities, machines, components, units, fleets, industrial assets, etc.
• the invention can in the future also be used to detect or predict other malfunctions such as specific and possibly critical physical or mental conditions of humans (illnesses, neurological malfunctions like epileptic seizures, genetic defects, etc.). Therefore, the component must be replaced by a part of the body, such as an organ or limb, and the unit would correspond to the human body.
• the single malfunctions would correspond to different specific conditions such as illnesses, neurologic malfunctions, genetic defects, etc.
• the parameters for detecting/predicting the parameters are described here with respect to malfunctions of technical entities, machines, components, units, fleets, industrial assets, etc.
• the invention can in the future also be used to detect or predict other malfunctions such as specific and possibly critical physical or mental conditions of humans (illnesses, neurological malfunctions like epileptic seizures, genetic defects, etc.). Therefore, the component must be replaced by a part of the body, such as an organ or limb, and the unit would correspond
• malfunction would be preferably parameters of the body, e.g. blood pressure, heart rate, temperature, etc.
• the invention can in the future also be used to determine the probability of certain achievements of the human body on the basis of the current parameters of the body. Therefore, the component must be replaced by the body or a part of the body.
• the single malfunctions would correspond to a single achievement. Achievements could be for example training goals, competition goals or diet goals. An example for a training goal could be to achieve a certain step frequency above a threshold over a certain period of time or a certain pulse frequency under a certain threshold. If the achievement comprises several sub-achievements, their probability can be combined as several single malfunctions can be combined to a general malfunction for technical entities.
• the parameters for predicting the achievement would be preferably parameters of the body, e.g. blood pressure, heart rate, step frequency, velocity during an activity, calorie consumption etc.
• a method for predicting a specific condition or achievement of a human body could comprise the steps of.
• the parameter of a human body comprises also parameters of his activity like a velocity, a step frequency, a pedalling rate, etc.

Landscapes

• Engineering & Computer Science (AREA)
• Business, Economics & Management (AREA)
• Physics & Mathematics (AREA)
• Economics (AREA)
• Entrepreneurship & Innovation (AREA)
• Human Resources & Organizations (AREA)
• Strategic Management (AREA)
• Theoretical Computer Science (AREA)
• General Physics & Mathematics (AREA)
• General Business, Economics & Management (AREA)
• Operations Research (AREA)
• Quality & Reliability (AREA)
• Tourism & Hospitality (AREA)
• Marketing (AREA)
• Game Theory and Decision Science (AREA)
• Educational Administration (AREA)
• Development Economics (AREA)
• Computational Mathematics (AREA)
• Mathematical Optimization (AREA)
• Artificial Intelligence (AREA)
• Probability & Statistics with Applications (AREA)
• Data Mining & Analysis (AREA)
• Evolutionary Computation (AREA)
• Mathematical Analysis (AREA)
• Algebra (AREA)
• Pure & Applied Mathematics (AREA)
• Computing Systems (AREA)
• General Engineering & Computer Science (AREA)
• Mathematical Physics (AREA)
• Software Systems (AREA)
• Testing And Monitoring For Control Systems (AREA)

Applications Claiming Priority (2)

Publications (1)

Family

ID=51421767

Family Applications (1)

Country Status (3)

Families Citing this family (19)

Family Cites Families (21)

• 2015
  • 2015-08-03 EP EP15747459.4A patent/EP3183698A1/de active Pending
  • 2015-08-03 US US15/504,725 patent/US10657453B2/en active Active
  • 2015-08-03 WO PCT/EP2015/067848 patent/WO2016026680A1/en active Application Filing

Non-Patent Citations (2)

Also Published As

Similar Documents

Legal Events